Skip to content

Notes · 3 min read

30Oct
2026
AI agentsB2B SaaS

AI agent MVP: approve the exact action, not just the conversation

A practical approval contract for agents that change records: preview, expiry, execution checks and recovery.

An assistant that drafts an answer and an agent that changes a customer record need different acceptance tests. Before adding a write action to an MVP, decide what a person is approving and how the application proves that the executed change matches that approval. This note is an editorial experiment: a proposed implementation checklist, not a report of measured customer results.

  1. 01

    Consider an internal operations agent preparing to move a delivery from Tuesday to Thursday. A useful approval preview identifies the order, the existing delivery date, the proposed date and any customer notification that will follow. A button labeled Approve should authorize that specific proposal. It should not silently authorize a later suggestion to change the address or send a different message. Keep the first release to one operational action so its consequences can be explained clearly.

  2. 02

    Separate proposal generation from execution. For client tools, Anthropic documents that the model returns a structured tool call and the application runs the operation. That gives application code a place to validate a proposed write before executing it. Tool arguments describe a request; they do not establish the caller's permissions. The application should check the signed-in operator's access to the target record and the permitted action independently of the model's wording.

  3. 03

    Store a proposal identifier with the exact target and proposed fields. Bind approval to that version of the proposal, the approving operator and an expiry chosen for the workflow. If the date or notification text changes, create a new proposal and show it again. If another operator has already changed the delivery, reject the stale proposal and ask for a fresh preview. An approval that was reasonable ten minutes ago may no longer describe the record being changed.

  4. 04

    Treat uncertain execution as a separate state. Suppose the scheduling service accepts the update but the response is lost. The agent should not announce failure and blindly repeat the write. Give each operation a stable identifier, record its outcome where possible, and reconcile the target system before retrying. When a downstream service cannot prevent duplicate requests, route uncertain outcomes to an operator instead of pretending that a retry is harmless. Show success only after the application has evidence of the resulting state.

  5. 05

    The tradeoff is speed versus review effort. Requiring confirmation for every read-only lookup makes an assistant tedious. Requiring a preview for a consequential write makes the initial pilot easier to inspect. Define the boundary per tool and effect: fetching availability can run immediately, while moving a delivery or notifying a customer may need approval. Do not let an instruction embedded in retrieved notes expand that boundary.

  6. 06

    Test the approval contract with concrete failures: edit the proposal after approval, revoke the operator's access, change the order in another session, double-click confirmation, and lose the execution response. A passing result preserves the approved scope, rejects stale or unauthorized writes, avoids duplicate effects where supported and exposes uncertain outcomes for recovery. Record proposal ID, operator reference, timestamps and outcome without copying unnecessary customer details into logs. These checks establish behavior; they do not prove the agent will increase conversions.

  7. 07

    Technical reference: Anthropic tool-use overview, https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview (reviewed 30 October 2026). The approval design and delivery example above are proposed application choices.

Need help with this?

Tell me about your project. I can help.

Send me a message about “AI agent MVP: approve the exact action, not just the conversation” or your own idea. Usually replies within one business day.

Keep exploring

Where to go next

Next step

Want help with your own project?

Send me a message. I will tell you what I would do first.